Determine whether additional resolution headroom improves LFPR

Determine whether increasing the Qwen3-VL resolution floor beyond the historically selected 4,194,304-pixel setting improves label-free precision refinement, using row-level predictions rescored against a common reference to resolve the inconsistent resolution-ladder measurements.

Background

The LFPR router uses a 4,194,304-pixel minimum image-processing budget. A staged pilot compared several resolution settings, but the archived summaries used inconsistent references: the 1,048,576- and 4,194,304-pixel results were reported relative to the default setting, whereas the 8,388,608-pixel result was reported relative to the 4,194,304-pixel arm.

Because these results cannot be placed on a common default-relative scale without rescoring the row-level predictions, the paper cannot establish whether performance peaked at 4,194,304 pixels or whether higher resolution would provide further gains. The authors explicitly leave the remaining resolution headroom unresolved.

References

Until the row-level predictions are rescored against one common reference, the defensible conclusion is only that $4{,}194{,}304$ was the historically selected operating point; remaining resolution headroom is unresolved.

— Where Grounding Accuracy Lives on the IoU Curve: Label-Free Inference-Time Boundary Refinement  (2608.19553 - Ma, 20 Aug 2026) in Appendix, Section 5.8, “Exploratory resolution checks do not establish a peak” (Section \ref{sec:headroom-exhausted})